Voice coding: dictate prompts for Claude Code and Cursor

"Voice coding" used to mean dictating open paren and semicolon and giving up after an hour. That is not what people mean now. With coding agents like Claude Code and Cursor, most of your work is writing instructions in plain language: what to change, why, what to leave alone. That is exactly what speech is good at. Here is how to dictate those prompts, what the built-in voice features do, and where a system-wide tool fills the gaps.

What changed: you now write more prose than code

When an agent writes the code, your job moves to the brief. A good brief is a paragraph or three: the bug, how to reproduce it, what you already tried, which files to avoid, how to test the result. Typed, that paragraph is a chore, so prompts get short and the agent guesses. Spoken, it takes half a minute and comes out fuller.

Voice is also a good fit for the other text developers write all day: commit messages, pull request descriptions, code review comments, tickets, and messages to teammates.

Built-in dictation in Claude Code

Claude Code has voice dictation in its command-line interface. According to its documentation, the essentials are:

• You turn it on with /voice. In the default hold mode you hold Space while you speak; /voice tap switches to tap-to-start, tap-to-send.
• Your speech appears live in the prompt, and you can mix typing and speaking in one message.
• It needs a claude.ai account and a local microphone. The audio is streamed to Anthropic's servers for transcription, and the docs say it is not processed locally. It does not work in cloud sessions or over SSH.
• The default dictation language is English. Russian and Polish are among the supported languages, and you set the language in /config.

It is a solid option if you live in the Claude Code terminal and are signed in with a claude.ai account. Details change quickly, so check the current docs page for your version.

Cursor

Cursor has its own voice input for the agent chat, and it also has a separate two-way voice chat. As of this writing I could not open Cursor's documentation from my research session, so I will not describe its shortcuts or limits here. Open the voice section in Cursor's official docs and settings to see what your version offers. Either way, such features usually work only inside the chat box of that one app.

Where built-in dictation stops

Built-in voice covers the prompt box. A developer's day has many more text fields:

• the terminal, where you write commit messages and shell notes
• the pull request page in the browser
• Slack, tickets and docs
• comments in the editor itself
• prompts to other assistants, such as the ones in talking to ChatGPT and Claude by voice

A system-wide dictation tool works in all of them with one key. If you already know you want that, the best dictation apps for Mac in 2026 compares the options.

A routine for dictating prompts on Mac

With Tiho the routine is the same in a terminal, Cursor or a browser tab:

1. Put the cursor in the prompt box of Claude Code, Cursor or any other field.
2. Hold right ⌥ and talk. Release, and the cleaned text is pasted at the cursor. For a long brief, double-tap ⌥ for hands-free, talk for a minute, then tap once to finish.
3. Use five beats. Goal, context, constraints, files or areas to avoid, and how to verify. For example: "The export button in the reports page crashes when the list is empty. Fix it. Keep the existing API unchanged. Don't touch the billing module. Add a test for the empty case and run the suite."
4. Correct by speaking. Say the sentence again if you misspoke. Filler words and abandoned starts are dropped.
5. Glance at names, then send. Package names, function names and acronyms are what any speech recognition gets wrong. Fix a spelling once and Tiho's dictionary remembers it next time.
6. Use snippets for the repeats. A short spoken phrase can expand into a paragraph you type all the time, such as "run the tests and show me the diff before committing".

Tiho does not turn speech into code symbols and is not a coding assistant. It is for the sentences around the code: prompts, messages and documentation.

Commit messages and review comments

Say what changed and why, in plain words: "Fix crash in the export button when the report list is empty. The empty state was never handled." Then adjust the first line by hand if your team uses a strict format. For review comments, speak the reasoning, which people usually skip when typing.

Languages and mixed sentences

Many developers think in Russian or Polish and write code in English. Tiho dictates English, Russian, Polish and Spanish, including mixed sentences with English terms inside, so "добавь тест для getUserById" or "dodaj test dla endpointu /orders" works without switching modes. See voice typing in 2026 for how it all fits together.

Privacy: your prompts contain your code

A spoken prompt can name internal projects, customers or secrets. Know where each tool sends audio. Claude Code's dictation sends it to Anthropic for transcription. Tiho has an on-device mode that works offline, and cloud processing happens only after you allow it; the privacy section explains what is stored. Never read secrets, keys or passwords aloud into any dictation tool.

FAQ

Can I dictate code itself, with brackets and semicolons?

That is a different job. Voice-coding tools that map speech to syntax exist, but for most people the practical win is dictating prompts and prose to a coding agent and letting the agent write the syntax.

Does Claude Code's /voice work over SSH or in the cloud?

No. The documentation says it needs a local microphone and does not work in cloud sessions or SSH sessions. A dictation tool on your own Mac types into the terminal window, so it works as long as you have a microphone and the window has focus.

Can I dictate in Russian or Polish?

Yes. Claude Code's dictation lists both among its supported languages, and Tiho dictates Russian, Polish and English, including mixed sentences.

What does it cost?

Tiho gives you 2,000 words a week free with no card. Paid plans have fair-use unlimited dictation; see the pricing section for current details.

Takeaway

When an agent writes the code, the prompt is the work, and spoken prompts are longer and clearer. Use built-in dictation if you only need it in one tool and you are happy with where it sends audio. Use a system-wide tool if you also want it for commits, reviews and messages. Try dictating your next five prompts and see how often the first answer is right.